A nonparametric copula approach to predict heart rate trajectories from single overnight wearable accelerometry

Wearable devices provide continuous, non-invasive monitoring of heart rate and motion at home, offering significant potential to advance precision health and medicine. Conventional heart rate prediction methods often depend on physiological signals through proprietary algorithms or rigid parametric forms. In this study, we explore a copula-based transformation method with a local log-quadratic likelihood kernel estimator and a nearest-neighbor bandwidth (Tll2nn) to estimate heart rate trajectories from only overnight triaxial accelerometer data. We test this approach on Apple Watch data from ~30 healthy participants during sleep. Our model effectively captures nonlinear dependencies between motion-derived features and heart rate fluctuations, outperforming a baseline prediction of constant mean heart rate in 20 of 26 subjects. We conclude that our methodology enables a novel, interpretable, computationally efficient and at-scale automation of heart rate trajectory tracking without relying on a physiological signal, even using data from a single overnight.Clinical Relevance- Automated tracking of heart rate through widely-utilized wearable devices can revolutionize remote patient monitoring. The technology enables early detection of physiological irregularities and improved management of sleep disorders and cardiovascular diseases, supporting timely intervention and enhanced health outcomes.